Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Adaptive Federated Learning to Optimize Integrated Flows in Cyber-Physical Data Centers

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An adaptive federated learning-to-optimization framework can solve day-ahead energy management for distributed data centers with a 0.08% cost error relative to centralized optimization, while keeping operational data private and verifiable.

desk verdict Good engineering, but the privacy story breaks at the label-generation stage. read the letter →

arxiv 2511.00623 v3 pith:V2KAED5G submitted 2025-11-01 eess.SY cs.SYmath.OC

Junhong Liu , Lanxin Du , Yujia Li , Rong-Peng Liu , Yunfeng Li , Fei Teng , Francis Yunhe Hou This is my paper
classification eess.SYcs.SYmath.OC
keywords datacentersfederatedlearninglearning-to-optimizationprivacypreservationmulti-energysystemsmixed-integerprogrammingverifiablesecretsharinghomomorphicencryption
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that coordinating electricity, heat, and data flows across geographically distributed data centers, normally a mixed-integer optimization requiring private data, can be decomposed into local neural-network prediction plus a small shared optimization, trained in a privacy-preserving way. If correct, data centers can cooperate on energy planning without exposing proprietary workload or cooling data, achieving near-centralized optimality at about one-ninth the compute time. The key is a learning-to-optimize pipeline with adaptive federated learning that tolerates heterogeneous local datasets, plus a cryptographic double-aggregation step that makes shared data both private and tamper-evident. The reported results show a 0.08% aggregate cost error against the centralized baseline and a solve-time reduction from ~5.30s to ~0.61s.

What carries the argument

The load-bearing mechanism is the learning-to-optimization decomposition: each data center runs an ensemble of multilayer perceptrons that predict its local decision variables from market prices and local inputs; only the first K weight blocks are shared in federated learning, masked by per-agent random weights and reconstructed via Lagrangian interpolation. An adaptive acceptance criterion, based on the median, slope, and standard deviation of recent loss trajectories relative to independent training, gates whether each agent adopts the federated update, using momentum mixing. At the optimization stage, the verifiable double aggregation scheme shares masked data (p_dc + w) and an encrypted

What would settle it

Re-run the experiments with training labels generated without solving the centralized problem—for example, from local subproblems that only use each data center's own data and fixed market prices—then measure the aggregate cost against the centralized MILP. If the relative error jumps well above 0.08%, the claim that the framework preserves near-optimality under privacy constraints is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a centralized mixed-integer linear program coupling power, heat, and data flows over multiple data centers can be replaced by a two-stage pipeline: each data center's local states (server counts, processing rates, temperatures, cooling power, queue lengths) are predicted by an ensemble of neural networks trained via adaptive federated learning, and only the residual system-level decisions (generator outputs, battery schedules, grid imports/exports) are solved as a small MILP. Training is made private through secret-shared aggregation with random masking and Lagrangian reconstruction; inference-time sharing of a few scalar variables per data center is made bo

Load-bearing premise

The method presumes that historical 'optimal decisions' for each data center are available as training labels, and for the coupled formulation such labels normally require the very centralized data collection and solving that the method aims to avoid.

Editorial extensions

If this is right

  • Data center operators can participate in day-ahead coordination without revealing workload, cooling, temperature, or queue-length time series to a central utility.
  • The pipeline runs in about 0.61 seconds per planning instance versus 5.30 seconds for centralized MILP, enabling faster what-if analysis or scaling to larger networks.
  • The adaptive acceptance criterion reduces negative transfer from heterogeneous local datasets, allowing agents with dissimilar data to selectively down-weight or skip federated updates.
  • The double-aggregation verification detects injected false data in the shared aggregate, so a utility can reject corrupted updates before solving the reduced optimization.
  • A near-zero aggregate cost error indicates that the learned surrogate introduces a negligible optimality gap under the tested operating conditions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The privacy boundary has a training-stage caveat the paper does not address: the neural networks are trained on historical 'optimal decisions,' which for the coupled problem generally require solving a centralized formulation with all data—so the privacy-preserving claim holds strictly only for the downstream inference and sharing stages.
  • The convergence guarantee is proved for an idealized convex, smooth objective satisfying a Polyak–Łojasiewicz condition; the actual neural-network training is nonconvex, so the theorem is indicative rather than directly applicable to the implemented system.
  • The same decomposition idea—learn local states from prices and local inputs, then optimize a small shared residual—could transfer to other multi-energy coordination problems, such as district heating or electric-vehicle fleet charging, wherever local states are predictable.
  • The acceptance criterion relies on hand-tuned thresholds and heuristics; a principled aggregation rule derived from the loss landscape could make the framework more robust and eliminate parameter tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes an adaptive federated learning-to-optimization framework for day-ahead energy management of geographically distributed data centers. A centralized multi-period mixed-integer program (P2) is approximated by locally trained neural networks that map prices, queues, and other inputs to data-center-level decisions; only a reduced problem (P3) for generators, storage, and grid trades is solved centrally, using masked aggregates of local variables. The method includes a privacy-preserving weighted aggregation protocol and a verifiable double-aggregation scheme for shared data, plus an acceptance criterion for federated model updates and an appendix convergence proof. Numerical experiments on a five-data-center test case report R² above 0.95 for learned variables, a 0.08% aggregate cost error relative to the centralized benchmark, and an approximately 8.7x runtime reduction.

Significance. If the claims are substantiated, the framework is a promising step toward distributed, privacy-aware data-center optimization: the numerical results on the presented test case are strong, the cryptographic aggregation is described concretely (CKKS parameters, verification tolerance), and the acceptance criterion is a practical mechanism for heterogeneous federated training. However, two load-bearing gaps prevent accepting the claims as stated. First, the training labels appear to require solving the centralized P2 with all data centers' private data, which breaks the end-to-end privacy boundary the paper emphasizes. Second, the appendix convergence proof analyzes an idealized SCAFFOLD-style algorithm with convex/PL assumptions that does not match Algorithm 1's plain neural-network training, and the proof contains a gradient-expectation inconsistency. The near-optimality evidence is also limited to a single test case.

major comments (4)
  1. [§III-A and §II-A] The training labels y=(x_i,t, r_i,t, η_eff_i,t, T_i,t, ...) are described as 'historical optimal decisions,' but in P2 these local variables are coupled across data centers by the system-wide power balance (1ad) and by generator ramping constraints (1w)-(1x) that tie time steps. Producing such labels therefore requires solving the centralized MILP P2 with access to all data centers' private data. The paper does not describe any private or distributed procedure for label generation, and the experiments (using dataset [23]) presumably generate labels by a central solve. Consequently, the privacy guarantee covers only the online inference/aggregation phase, not the end-to-end workflow. This directly affects the abstract claim 'preserves the privacy and integrity of shared data.' Please provide a distributed/private label-generation method, or explicitly restrict the privacy claim and reposi
  2. [Appendix, Eqs. (16)-(21), (35a)] The convergence theorem assumes the update θ_{r,i}=θ̄_r−η_loc_r·g̃_{r,i} with global-unbiased control variates satisfying (19a)-(19c), and assumes f is convex and L-smooth plus a PL condition. Algorithm 1 in §III-A instead trains an MLP with a composite MSE/Huber loss via ordinary local steps; no control variates are introduced, and the neural-network objective is nonconvex. Additionally, (19a) defines E[g̃_i|θ̄_r]=∇f(θ̄_r), while the derivation of (26) uses E[g̃_i]=∇f_i(θ̄_r) for the local functions; these are inconsistent. As written, the O(1/√T) guarantee does not apply to the implemented algorithm. Please align the theory with the algorithm, state clearly that the theorem is for an idealized scheme, and fix the gradient-expectation inconsistency.
  3. [§IV-B, Table II] The near-optimality claim rests on a single 24-hour test instance. Table II reports one set of component costs; the ten runs in §IV-C are only for timing. The paper should report results over multiple days/market/weather conditions with means and spreads. As presented, a single 0.08% aggregate cost error is not sufficient to support the general claim of near-optimality across heterogeneous data-center operating regimes.
  4. [§III-B, P3] Learned local variables (including p_dc_i,t and δ_i,t) are inputs to P3, but no mechanism enforces the original physical constraints for these variables: queue dynamics (1c), temperature dynamics (1p)-(1u), and integer server commitment (1i). If the learned values violate these constraints, the P3 solution is not necessarily feasible for the original P2, so the cost comparison to M0 is not a valid optimality gap. The paper does not report constraint-violation rates or a feasibility-repair step. This is central to the near-optimality claim and should be addressed explicitly.
minor comments (6)
  1. [§IV-C (before Fig. 8)] The text contains apparent font-encoding artifacts (strings such as '/uni00000013/uni00000011/...'). These should be corrected before publication.
  2. [§II-A, P2] The paper calls P2 a MILP, but the objective (3a) still contains the quadratic term β_g·p_g,t². Unless β_g=0, P2 is a mixed-integer quadratic program, not a MILP. Please correct the description.
  3. [Eq. (15a)] The NRMSE formula uses N and n in a confusing way; please define the sample-count symbols consistently. The formula as printed appears to mix an outer average over N with an inner root-mean-square over n.
  4. [V, Conclusions] Typo: 'flows flows' should be 'flows'.
  5. [Table II] For Cost 3 (SLA penalty), the base value in M0 is only 1.84 dollars, so relative errors of 106.52% and 49.46% are misleading. Report absolute deviations for tiny-cost components alongside relative ones.
  6. [Reference [23]] Please include an access/version date and, if possible, a DOI for the GitHub dataset.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the 'near-optimal' benchmark is the same centralized P2 that supplied the training labels; privacy/FL claims remain independent.

  1. fitted input called prediction [Section III-A (expert demonstrations) and Section IV-B (Table II, M0 benchmark)]
    "Historical decisions perform as the expert demonstrations, informing the learning process of the agent for each data center. ... M0 denotes the reference centralized optimization ... Error 1−3 correspond to the relative errors obtained from methods M 1−3, respectively, evaluated with respect to results from method M 0."

    The local-decision network (Eq. (4)) is fitted to the historical optimal decisions y of P2 via the training loss in Algorithm 1 (MSE/Huber). Table II then evaluates the learned pipeline's cost against M0, which is the same P2 that produced those labels. If the network reproduced y exactly, P3 (fixing those local decisions and re-optimizing only X†) would return M0's cost by construction; the 0.08% aggregate error is therefore a measure of how well the model imitated the centralized oracle, not an independent test of optimality. The near-optimality claim is thus partially a report of fit quality, although prediction error can be nonzero, so the reduction is not a strict identity.

full rationale

The derivation chain is otherwise self-contained: P2 is a standard MILP formulation (Sec. II-A), P3 is obtained by fixing the learned local variables (Sec. III-B), and the double-aggregation privacy mechanism is analyzed with its own verification equation. The convergence proof in the Appendix is a standard federated-averaging argument (Eqs. (16)-(35)) using convexity, L-smoothness and the PL condition, with the acceptance criterion entering only as a bounded coefficient C_{r,i} in [0,1]; it does not assume the conclusion. Self-citations [7], [18], [20], [23] are used for background, crypto primitives, countermeasures, and data, and none supplies a uniqueness or optimality theorem that forces the main result. A separate deployment gap should be noted but is not circularity per se: producing 'historical optimal decisions' for the labels appears to require solving P2 with all data centers' data (constraints such as (1ad) couple all local variables), so the privacy claim is scoped to online inference and aggregation unless a private label-generation method is supplied. Weighing all of this, the only meaningful circularity is the benchmark-dependence of the near-optimality claim, which is partial rather than a by-construction reduction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claims rest on standard federated-learning assumptions that do not hold for the actual neural network training, plus a bootstrapping assumption about training labels that is not resolved. The hand-tuned acceptance thresholds and unreported ensemble size are additional free choices.

free parameters (4)
  • Acceptance criterion thresholds and weights = τ1=0.1, τ2=0.01, τ3=0.8, κ1=0.4, κ2=0.3, κ3=0.3, window L=20
    Set by hand in Algorithm 1 with no sensitivity analysis; they determine which model updates are accepted, so the adaptive behavior and resulting near-optimality depend directly on these values.
  • Momentum and training schedule = β=0.9, warm-up 1500 steps, 800 local steps/round, 25 rounds, batch size 512, dropout 0.15, hidden width [256,1024,256]
    Chosen in Section IV-A without justification; the reported R² and cost accuracy are contingent on these hyperparameters.
  • Ensemble size N = not reported
    The ensemble averaging (Eq. (12)) is central to the method's predictions and uncertainty estimates, but the number of networks N is never given in the simulation setup, making the exact algorithm unreproducible.
  • Verification tolerance ψ = ψ=10^-4
    Introduced ad hoc in Section III-C to absorb encryption/decryption noise; it also determines the sensitivity of false-data-injection detection.
assumptions (6)
  • domain assumption Local losses f_i are convex and L-smooth
    Used in Appendix A before Eq. (16) to derive the descent lemma and convergence bound; ReLU neural network losses are non-convex and not globally smooth, so the proof does not apply to the actual trained models.
  • domain assumption Global objective satisfies the Polyak-Łojasiewicz condition
    Invoked after Eq. (31) to get linear convergence; no evidence is given that the federated loss satisfies PL.
  • ad hoc to paper Bounded data heterogeneity and gradient noise with control variates
    Assumptions (19a)-(21) assume SCAFFOLD-style global-unbiased control variates and bounded heterogeneity ζ, but Algorithm 1 does not implement control variates; the convergence analysis does not model the actual residual gradient bias.
  • domain assumption Historical optimal decisions are available as training labels
    Section III-A: 'Historical decisions perform as the expert demonstrations.' Since optimal local decisions in P2 depend on global coupling constraints, obtaining these labels in a privacy-preserving setting is unresolved.
  • domain assumption Honest-but-curious utility and secure communication channels
    The secret-sharing step requires secure channels ('which can be realized via the CRT-Paillier or CKKS cryptosystems'), and the integrity verification assumes the utility broadcasts correct intermediate values; a malicious utility is not considered.
  • domain assumption Prediction errors do not break feasibility of P3
    P3 fixes data-center variables at predicted values; if predictions violate coupling constraints such as temperature dynamics or SLA constraints, the simplified problem could be infeasible or far from optimal. No suboptimality bound as a function of prediction error is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Federated Learning to Optimize Integrated Flows in Cyber-Physical Data Centers." pith.science (2026). https://pith.science/paper/V2KAED5G

@misc{pith2026251100623,
  author       = {Pith},
  title        = {Pith review of: Adaptive Federated Learning to Optimize Integrated Flows in Cyber-Physical Data Centers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2KAED5G}},
  note         = {Machine review of arXiv:2511.00623}
}
read the original abstract

Data centers play an increasingly critical role in societal digitalization, yet their rapidly growing energy demand poses significant challenges for sustainable operation. To enhance the energy efficiency of geographically distributed data centers, this paper formulates a multi-period optimization model that captures the interdependence of electricity, heat, and data flows. The optimization of such integrated multi-domain flows inherently involves mixed-integer formulations and the access to proprietary or sensitive datasets, which correspondingly exacerbate computational complexity and raise data-privacy concerns. To address these challenges, an adaptive federated learning-to-optimization approach is proposed, accounting for the heterogeneity of datasets across distributed data centers. To safeguard privacy, cryptography techniques are leveraged in both the learning and optimization processes. A model acceptance criterion with convergence guarantee is developed to improve learning performance and filter out potentially contaminated data, while a verifiable double aggregation mechanism is further proposed to simultaneously ensure privacy and integrity of shared data during optimization. Theoretical analysis and numerical simulations demonstrate that the proposed approach preserves the privacy and integrity of shared data, achieves near-optimal performance, and exhibits high computational efficiency, making it suitable for large-scale data center optimization under privacy constraints.

Figures

Figures reproduced from arXiv: 2511.00623 by the authors.

Figure 1
Figure 1. Scheme of the data center energy management. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Personalized adaptive federated learning (offline phase). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Personalized adaptive federated learning (online phase). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Verifiable data sharing via double aggregation mechanism. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Training results for the adaptive federated learning. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Prediction performance for data center 5. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: System states from the adaptive federated learning-to-optimization. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Time consumption for different approaches. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 3 linked inside Pith

  1. [23]

    data-center-dataset,

    J. Liu, R.-P. Liu, Y . Hou, and etc., “data-center-dataset,” https://github. com/johnny-eee/data-center-dataset, 2025

    2025

  2. [1]

    Research and technolo- gies for next-generation high-temperature data centers–state-of-the-arts and future perspectives,

    Y . Zhang, K. Shan, X. Li, H. Li, and S. Wang, “Research and technolo- gies for next-generation high-temperature data centers–state-of-the-arts and future perspectives,”Renewable and Sustainable Energy Reviews, vol. 171, p. 112991, 2023

    2023

  3. [2]

    Usage impact on data center electricity needs: A system dynamic forecasting model,

    M. Koot and F. Wijnhoven, “Usage impact on data center electricity needs: A system dynamic forecasting model,”Applied Energy, vol. 291, p. 116798, 2021

    2021

  4. [3]

    A highly scalable network architecture for optical data centers,

    W. Fan, Y . Pan, F. Xiao, P. Zhang, L. Han, and S.-Y . Hsieh, “A highly scalable network architecture for optical data centers,”IEEE Transactions on Computers, 2025

    2025

  5. [4]

    Exploiting internet data centers as energy prosumers in integrated electricity-heat system,

    X. Yin, C. Ye, Y . Ding, and Y . Song, “Exploiting internet data centers as energy prosumers in integrated electricity-heat system,”IEEE Trans- actions on Smart Grid, vol. 14, no. 1, pp. 167–182, 2022

    2022

  6. [5]

    Optimizing energy consumption for data centers,

    H. Rong, H. Zhang, S. Xiao, C. Li, and C. Hu, “Optimizing energy consumption for data centers,”Renewable and Sustainable Energy Reviews, vol. 58, pp. 674–691, 2016

    2016

  7. [6]

    Collaborative response of data center coupled with hydrogen storage system for renewable energy absorption,

    X. Long, Y . Li, Y . Li, L. Ge, H. B. Gooi, C. Chung, and Z. Zeng, “Collaborative response of data center coupled with hydrogen storage system for renewable energy absorption,”IEEE Transactions on Sus- tainable Energy, vol. 15, no. 2, pp. 986–1000, 2023

    2023

  8. [7]

    Synergising hierarchical data centers and power networks: A privacy-preserving approach,

    J. Liu, F. Teng, and Y . Hou, “Synergising hierarchical data centers and power networks: A privacy-preserving approach,”IEEE Transactions on Smart Grid, pp. 1–1, 2025

    2025

Show all 31 references
  1. [8]

    Cyberse- curity in smart grids, challenges and solutions,

    D. Faquir, N. Chouliaras, V . Sofia, K. Olga, and L. Maglaras, “Cyberse- curity in smart grids, challenges and solutions,”AIMS Electronics and Electrical Engineering, vol. 5, no. 1, pp. 24–37, 2021

    2021

  2. [9]

    The challenges of iot addressing security, ethics, privacy, and laws,

    A. Karale, “The challenges of iot addressing security, ethics, privacy, and laws,”Internet of Things, vol. 15, p. 100420, 2021

    2021

  3. [10]

    Protecting data in the 21st century: Challenges, strategies and future prospects,

    A. K. Saraswat and V . Meel, “Protecting data in the 21st century: Challenges, strategies and future prospects,”Information technology in industry, vol. 10, no. 2, pp. 26–35, 2022

    2022

  4. [11]

    Privacy-preserving edge fed- erated learning for intelligent mobile-health systems,

    A. Aminifar, M. Shokri, and A. Aminifar, “Privacy-preserving edge fed- erated learning for intelligent mobile-health systems,”Future Generation Computer Systems, vol. 161, pp. 625–637, 2024

    2024

  5. [12]

    Federated learning for real-time demand response by data centers toward energy efficiency and privacy preservation,

    M. Alhazmi, A. P. Zhao, W. Li, and C. Yang, “Federated learning for real-time demand response by data centers toward energy efficiency and privacy preservation,”IEEE Access, 2025

    2025

  6. [13]

    Federated learning for 6g: Applications, challenges, and opportunities,

    Z. Yang, M. Chen, K.-K. Wong, H. V . Poor, and S. Cui, “Federated learning for 6g: Applications, challenges, and opportunities,”Engineer- ing, vol. 8, pp. 33–41, 2022

    2022

  7. [14]

    A lightweight mini-batch federated learning approach for attack detection in iot,

    M. S. Ahmad and S. M. Shah, “A lightweight mini-batch federated learning approach for attack detection in iot,”Internet of Things, vol. 25, p. 101088, 2024

    2024

  8. [15]

    Federated learning approach for secured medical recommendation in internet of medical things using homomor- phic encryption,

    E. A. Mantey, C. Zhou, J. H. Anajemba, J. K. Arthur, Y . Hamid, A. Chowhan, and O. O. Otuu, “Federated learning approach for secured medical recommendation in internet of medical things using homomor- phic encryption,”IEEE Journal of Biomedical and Health Informatics, vol. 28,...

    2024

  9. [16]

    An advanced data fabric architecture leveraging homomorphic encryption and federated learning,

    S. A. Rieyan, M. R. K. News, A. M. Rahman, S. A. Khan, S. T. J. Zaarif, M. G. R. Alam, M. M. Hassan, M. Ianni, and G. Fortino, “An advanced data fabric architecture leveraging homomorphic encryption and federated learning,”Information Fusion, vol. 102, p. 102004, 2024

    2024

  10. [17]

    Towards carbon-aware container orchestration: Predicting workload energy consumption with federated learning,

    Z. Saad, J. Yang, H. Leung, and S. Drew, “Towards carbon-aware container orchestration: Predicting workload energy consumption with federated learning,”arXiv preprint arXiv:2510.03970, 2025

    2025

  11. [18]

    Privacy- preserving peer-to-peer energy trading via hybrid secure computations,

    J. Liu, Q. Long, R.-P. Liu, W. Liu, X. Cui, and Y . Hou, “Privacy- preserving peer-to-peer energy trading via hybrid secure computations,” IEEE Transactions on Smart Grid, vol. 15, no. 2, pp. 1951–1964, 2023

    1951

  12. [19]

    Tenseal: A library for encrypted tensor operations using homomorphic encryption,

    A. Benaissa, B. Retiat, B. Cebere, and A. E. Belfedhal, “Tenseal: A library for encrypted tensor operations using homomorphic encryption,” arXiv preprint arXiv:2104.03152, 2021

    2021 arXiv

  13. [20]

    Byzantine-resilient distributed p2p energy trading via spatial-temporal anomaly detection,

    J. Liu, Q. Long, R.-P. Liu, W. Liu, and Y . Hou, “Byzantine-resilient distributed p2p energy trading via spatial-temporal anomaly detection,” IEEE Transactions on Smart Grid, 2025

    2025

  14. [21]

    Blockchain cloud computing: comparative study on ddos, mitm and sql injection attack,

    N. Singh, H. P. Singh, A. Mishra, A. Khare, M. Swarnkar, and S. K. Almas, “Blockchain cloud computing: comparative study on ddos, mitm and sql injection attack,” in2024 IEEE International Conference on Big Data & Machine Learning (ICBDML). IEEE, 2024, pp. 73–78

    2024

  15. [22]

    Gurobi Optimizer Reference Manual,

    Gurobi Optimization, LLC, “Gurobi Optimizer Reference Manual,”

  16. [24]

    Scaffold: Stochastic controlled averaging for federated learn- ing,

    S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “Scaffold: Stochastic controlled averaging for federated learn- ing,” inInternational conference on machine learning. PMLR, 2020, pp. 5132–5143

    2020

  17. [25]

    Local sgd converges fast and communicates little,

    S. U. Stich, “Local sgd converges fast and communicates little,”arXiv preprint arXiv:1805.09767, 2018

    2018 arXiv

  18. [26]

    Federated learning based on dynamic regularization,

    D. A. E. Acar, Y . Zhao, R. M. Navarro, M. Mattina, P. N. Whatmough, and V . Saligrama, “Federated learning based on dynamic regularization,” arXiv preprint arXiv:2111.04263, 2021

    2021 arXiv

  19. [27]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,”Proceedings of Machine learning and systems, vol. 2, pp. 429–450, 2020

    2020

  20. [28]

    S. P. Boyd and L. Vandenberghe,Convex optimization. Cambridge university press, 2004

    2004

  21. [29]

    On the linear speedup analysis of communi- cation efficient momentum sgd for distributed non-convex optimization,

    H. Yu, R. Jin, and S. Yang, “On the linear speedup analysis of communi- cation efficient momentum sgd for distributed non-convex optimization,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 7184–7193

    2019

  22. [30]

    Linear convergence of gradient and proximal-gradient methods under the polyak-łojasiewicz condition,

    H. Karimi, J. Nutini, and M. Schmidt, “Linear convergence of gradient and proximal-gradient methods under the polyak-łojasiewicz condition,” inJoint European conference on machine learning and knowledge discovery in databases. Springer, 2016, pp. 795–811

    2016

  23. [2024]

    Available: https://www.gurobi.com

    [Online]. Available: https://www.gurobi.com

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.